Add benches for encrypt/decrypt roundtrip

[nyonson]

Using cargo's builtin bench tool with the unstable test feature (so requires nightly toolchain). It has been around for few years now so I don't think a big risk, and rather not introduce a dependency.

I used the benches to run an experiment with the fschacha20poly1305 module. I thought performance could be improved by using a shared buffer in between calls (e.g. nonce() and rekey()) instead of having to zero-allocate arrays each time. But this actually resulted in a small performance decrease, I guess probably something to do with caching and optimizations around the small, stack-based buffers.